Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarScience Daily AI→ original

An AI Model Reads Routine ECGs to Find a Group Where 7% Die Suddenly Each Year, Versus 4.6%

More than 300,000 people die from sudden cardiac arrest in the US each year, and doctors have a poor sense of who needs a defibrillator in advance. A neural network from UC Berkeley, trained on 440,000 Swedish ECGs, identified a group in which 7% of patients die within a year. The conventional criterion, how much blood the heart pumps per beat, yields a group with 4.6%. Running the model requires only routine ECGs from hospital archives, with no new test needed. Checks in San Diego and Taipei showed the signal works beyond Sweden. About two-thirds of implanted defibrillators never fire. Assembling the three datasets took roughly 10 years.

AI-processed from Science Daily AI; edited by Hamidun News
An AI Model Reads Routine ECGs to Find a Group Where 7% Die Suddenly Each Year, Versus 4.6%
Source: Science Daily AI. Collage: Hamidun News.
Listen to article

Imagine an alarm system that is supposed to flag users who are about to leave, but it catches the wrong people and misses the most important cases. That is roughly how medicine handles sudden cardiac arrest today. A treatment exists: an implantable defibrillator stops a dangerous rhythm failure with an electrical shock.

But it is nearly impossible to know in advance who needs one. Researchers at the University of California, Berkeley trained a neural network to look for clues in a routine ECG and published the results in Nature. The group the AI selected turned out to be higher risk: 7% of people in it die of sudden cardiac arrest each year.

In the group selected by the standard physician criterion (how well the heart pumps blood with each contraction), that figure is 4.6%. In other words, the AI pointed more accurately to the people genuinely facing danger.

Why current screening falls short

More than 300,000 people die from sudden cardiac arrest in the US each year. This is not a heart attack: in a heart attack a vessel becomes blocked, whereas here the electrical system that sets the heartbeat fails. About 90% of people whose arrest happens outside a hospital die within a few minutes.

Today, the decision to implant a defibrillator mostly rests on the volume of blood the heart ejects per beat. If the pump falls below a threshold, the patient is considered a candidate for an implant. The method has two weak points. First, the measurement requires a separate test, and many future victims never take it because they have no reason to be worried. Second, roughly two-thirds of high-risk patients with an implant never have the device fire. Some people undergo an expensive invasive procedure with no benefit, while others die without warning.

How the neural network was trained

Ziad Obermeyer's team started with six years of ECGs from Sweden's unified healthcare system: more than 440,000 recordings. They matched these to death certificates. The model thus saw three groups: healthy people, people with known cardiac risks, and those who later died suddenly from a cardiac cause. It learned on its own to find combinations of peaks, waves, and other signal features associated with elevated risk.

The algorithm was then tested on two independent datasets: two years of de-identified ECGs from the San Diego health system and data from Taipei. This matters because a model trained on Swedes could simply have memorized the quirks of one population. The checks in the US and Taiwan showed that the signal it found also works in other countries.

The algorithm also identified a broader risk group: thousands of people who current guidelines would consider low risk. In addition, it needs only the routine ECGs that hospitals already collect, so no new type of test is required.

What comes next, and the limitations

Routine clinical use is still some way off. The authors are working with medical systems in Sweden, Taiwan, and the US to run the algorithm across hospital ECG databases. The plan: doctors will contact patients who receive a high score and offer them a patch that continuously records cardiac activity. That data should help clarify which processes in the heart produce the signal, and for some patients it may become grounds for defibrillator implantation.

There is also an honest caveat: scientists do not yet know what in the heart changes the electrical pattern the model learned. This is a powerful practical tool, but it does not yet come with an explanation. Clinical deployment will require prospective validation and agreement with regulators.

For the medical AI industry, there are two lessons. The first: value often comes not from the model architecture but from access to data. Obermeyer says such data is "incredibly hard to get." Assembling the three datasets took him about ten years, alongside the Dandelion Health and Nightingale Open Science projects he helped create. The second lesson: the cheapest and most familiar medical signals, such as the ECG, may hold more information than a physician can see. Teams building clinical AI should look for applications in archives that already exist and must test their models on data from other countries and hospitals.

Story

Key Questions Answered

Why current screening falls short?

More than 300,000 people die from sudden cardiac arrest in the US each year. This is not a heart attack: in a heart attack a vessel becomes blocked, whereas here the electrical system that sets the heartbeat fails. About 90% of people whose arrest happens outside a hospital die within a few minutes. Today, the decision to implant a defibrillator mostly rests on the volume of blood the heart ejects per beat. If the pump falls below a threshold, the patient is considered a candidate for an implant.

How the neural network was trained?

Ziad Obermeyer's team started with six years of ECGs from Sweden's unified healthcare system: more than 440,000 recordings. They matched these to death certificates. The model thus saw three groups: healthy people, people with known cardiac risks, and those who later died suddenly from a cardiac cause. It learned on its own to find combinations of peaks, waves, and other signal features associated with elevated risk.

What comes next, and the limitations?

Routine clinical use is still some way off. The authors are working with medical systems in Sweden, Taiwan, and the US to run the algorithm across hospital ECG databases. The plan: doctors will contact patients who receive a high score and offer them a patch that continuously records cardiac activity. That data should help clarify which processes in the heart produce the signal, and for some patients it may become grounds for defibrillator implantation.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…